Ruth Etzioni

PhD

Fred Hutchinson Cancer Research Center

Full Member

PLCO (Learn more about this study)

PLCO-9

Dec 11, 2012

Reconciling results of PSA screening efficacy using PLCO and ERSPC data: CISNET modeling project

The Cancer Intervention and Surveillance Modeling Network (CISNET) is a consortium of NCI-funded investigators using modeling techniques to study the impacts of cancer control interventions on national trends in cancer rates. Following publication of apparently conflicting results about PSA efficacy from the PLCO and the ERSPC, the CISNET prostate team initiated collaboration with investigators from both trials to determine whether there is a range of values for PSA screening efficacy that is consistent with both trials. Modeling provides a unique opportunity to systematically control for differences in trial design and implementation (pre-trial screening, screening ages, screen attendance, inter-screening intervals, thresholds for biopsy referral, control arm screening rates), background clinical practice, and receipt of cancer therapies. Using three independently developed CISNET models enhances the reliability of modeling results and provides a sensitivity analysis on model structures and underlying assumptions.

Based on agreements with trial investigators, trial data were released and studied in phases. Using trial data on screening and incidence for the intervention groups with up to 7 years of follow-up, the CISNET models successfully replicated testing patterns, follow-up distributions, and numbers of observed prostate cancers by age, year, stage, and grade in both trials, and the CISNET team presented a progress report to trial scientific committees in April 2011. In October 2012 we received screening and incidence information for trial control groups with up to 6 years of follow-up and are currently incorporating these data in our models.

The proposed project continues the original plan to apply the CISNET models to examine whether the trials provide consistent evidence regarding PSA screening efficacy. To do so, we are making a Standard Request for access to the Prostate and the Prostate Treatments datasets. As in the original proposal (PLCO Study ID 2009-00182), we will use comparable data and undertake the same modeling steps for both trials. Specifically, we will recalibrate the CISNET models’ natural history and clinical practice pattern parameters as necessary to reproduce incidence patterns by age, year, stage, and grade for both the intervention and control groups. We will differentiate between screen-detected and interval cases and estimate lead time for screen-detected cases. Using a common model of prostate cancer survival, we will estimate PSA efficacy as a (linear or non-linear) single-parameter function of lead time. Means and variances of this parameter will be estimated, allowing for comparison of the parameter value between trials to evaluate the hypothesis of a single underlying value of the parameter. We will report ranges of values of PSA efficacy consistent with each trial given its distribution of lead times, examine overlap of these ranges between the trials, and compare results across models. In addition, the final models will be used to simulate the two trials under perfect compliance and zero contamination, and we will quantify the effects of differences in trial design and implementation, background clinical practice, and receipt of cancer therapies between the trials in explaining the discrepancy in the trial mortality rate ratios with up to 13 years of follow-up.

1. To estimate, using standardized methods and models, screening frequencies in
both arms of each trial: PLCO and ERSPC.

2. To fit/calibrate the three CISNET models to prostate cancer incidence in the control and intervention arms of each trial. Common input on PSA screening frequencies will be used. Re-calibration of the natural history parameters will be considered and documented. The models will attempt to fit the incidence data in each arm of each trial by time period, age, method of detection (screen detected or not), stage and grade. Specific questions being addressed include: How different do the natural history parameters need to be to reproduce cancer incidence in the PLCO and ERSPC trials?
How much higher is incidence in the control arm of the PLCO as compared to the overall US?

3. To estimate initial treatment practices in both arms of each trial in a consistent manner. Since part of the decline in prostate cancer mortality can be attributed to improvements in prostate cancer treatment, it is important to estimate and
compare treatment practices in each trial.

4. To fit, using the cancer incidence and treatment modeling results, the mortality data for each trial and specifically, to estimate the mean and variance of a screening survival benefit parameter for each trial. Each CISNET model will be
restructured so that improvement in post-lead time survival due to screening is modeled in terms of a single survival benefit parameter (e.g., a hazard ratio or a relative cure probability).

5. Under the range of survival benefit parameter values consistent with the trial findings (e.g., 95% confidence interval), the models will project the benefits that would be expected when comparing a group screened according to a given protocol with 100% compliance to a completely unscreened group. This aim is therefore to project observed results to an uncontaminated version of the trials.